Abnormal account identification method and related device

Abstract

The invention relates to the technical field of computers, and provides an abnormal account recognition method and a related device, which are used for improving the abnormal account recognition accuracy, and the method comprises the steps: obtaining a complaint text generated for a to-be-detected account in a set collection period; on the basis of a target industry category to which the to-be-detected account belongs, a target recognition model is adopted to perform abnormal behavior analysis on the complaint text to obtain a corresponding analysis result, and the target recognition model is obtained by utilizing each historical complaint text through iterative training; and based on the obtained analysis result, identifying the to-be-detected account number, and obtaining an account number identification result of the to-be-detected account number. In the embodiment of the invention, the artificial intelligence technology is utilized to perform abnormal behavior analysis on the complaint text, and the adaptability and sensitivity of complaint texts of different industries are enhanced in combination with the target industry category to which the to-be-detected account belongs, so that the recognition accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology and provides a method for identifying abnormal accounts and related devices. Background Art

[0002] With the development of Internet technology and e-commerce, a large number of e-commerce platforms have emerged. Consumers can purchase products through online payment and enjoy the convenience of shopping on e-commerce platforms without leaving home.

[0003] At present, there are a large number of accounts connected to e-commerce platforms as product providers, and the quality of the accounts varies greatly. Although normal accounts generally account for the vast majority, there are also some abnormal accounts hidden among them, such as gambling accounts, fraud accounts, accounts with transaction disputes, etc. Such accounts will cause huge property losses to consumers.

[0004] In related technologies, in order to control risks, simple transaction limit controls are usually performed based on the industry category of the account. For example, accounts that trade physical items such as furniture are given larger transaction limits, while accounts that trade virtual items such as game props are given smaller transaction limits.

[0005] However, the above-mentioned transaction limit control is only based on the industry category of the account. It has a certain effect on normal accounts that comply with the law, but it is difficult to effectively control risks for abnormal accounts. For example, abnormal accounts can avoid transaction limit control by registering industry categories with larger transaction limits. Summary of the Invention

[0006] The embodiments of the present application provide a method and related apparatus for identifying abnormal accounts, so as to improve the accuracy of identifying abnormal accounts.

[0007] In a first aspect, an embodiment of the present application provides a method for identifying abnormal accounts, comprising:

[0008] Obtain at least one complaint text generated for the account to be tested within the set collection period;

[0009] Based on the target industry category to which the account to be detected belongs, a target recognition model is used to perform abnormal behavior analysis on each of the at least one complaint texts to obtain corresponding analysis results, wherein the target recognition model is obtained through iterative training using various historical complaint texts;

[0010] Based on the at least one analysis result obtained, the account to be detected is identified to obtain an account identification result of the account to be detected.

[0011] In a second aspect, an embodiment of the present application provides a device, including:

[0012] A text acquisition unit, configured to acquire at least one complaint text generated for the account to be detected within a set collection period;

[0013] a text analysis unit configured to perform abnormal behavior analysis on each of the at least one complaint texts based on the target industry category to which the account to be detected belongs, using a target recognition model to obtain corresponding analysis results, wherein the target recognition model is obtained through iterative training using various historical complaint texts;

[0014] The account analysis unit is configured to identify the account to be detected based on at least one analysis result obtained, and obtain an account identification result of the account to be detected.

[0015] In one possible implementation, the text analysis unit is further configured to:

[0016] Based on each historical complaint text, a training sample set is constructed. Each training sample carries corresponding label data. The label data is used to characterize the multiple abnormal behavior categories to which the corresponding training sample belongs. There is a hierarchical relationship between the multiple abnormal behavior categories.

[0017] Based on the training sample set, the initial recognition model is iteratively trained to obtain a trained initial recognition model, and based on the trained initial recognition model, the target recognition model is constructed.

[0018] In one possible implementation, the initial recognition model includes: multiple recognition sub-models, each recognition sub-model is used to predict a level of abnormal behavior category, and during each iteration, the text analysis unit is used to perform the following operations:

[0019] Obtaining a batch of training samples, and using the multiple recognition sub-models to perform label prediction on the training samples, to obtain prediction results corresponding to the training samples at multiple levels;

[0020] Based on the label data carried by each of the training samples and in combination with the multiple prediction results corresponding to each of the training samples, a model loss is obtained, and model parameters are adjusted based on the model loss.

[0021] In one possible implementation, when using the multiple recognition sub-models to perform label prediction on each training sample and obtain prediction results corresponding to each training sample at multiple levels, the text analysis unit is specifically configured to:

[0022] For multiple recognition sub-models, perform the following operations respectively:

[0023] Using a feature extraction network in a recognition sub-model, feature extraction is performed on each of the training samples to obtain corresponding text features;

[0024] The classification network in the recognition sub-model is used to classify the obtained text features respectively, and obtain the prediction results corresponding to the training samples at the corresponding levels.

[0025] In one possible implementation, when constructing a training sample set based on the historical complaint texts, the text analysis unit is specifically configured to:

[0026] Based on the historical complaint texts and their corresponding industry type texts, construct initial samples, each initial sample including a historical complaint text and its corresponding industry type text;

[0027] According to the set data enhancement method, data enhancement is performed on the initial samples to obtain the expanded samples, and the training sample set is obtained based on the initial samples and the expanded samples.

[0028] In one possible implementation, when performing data enhancement on the initial samples according to the set data enhancement method to obtain the expanded samples, the text analysis unit is specifically configured to:

[0029] For at least one of the initial samples, perform at least one of the following operations:

[0030] Replacing at least one word contained in a historical complaint text in an initial sample with a synonym of the at least one word to obtain a corresponding expanded sample;

[0031] Replacing keywords contained in a historical complaint text in an initial sample with interference words similar in font to the keywords to obtain a corresponding expanded sample;

[0032] Based on the semantic information of the historical complaint text of an initial sample, similar text expansion is performed on an initial sample to obtain a corresponding expanded sample.

[0033] In a possible implementation, the text analysis unit is further configured to generate label data for a sample data in the following manner:

[0034] Using a label annotation model, a training sample is labeled multiple times to obtain labeling results corresponding to each of the multiple label annotations, each labeling result representing the abnormal behavior category to which the training sample belongs;

[0035] Results are aggregated based on the multiple labeling results obtained to obtain a soft label for the training sample, and the soft label is used as label data for the training sample, wherein the soft label includes a label probability that the training sample belongs to each abnormal behavior category.

[0036] In one possible implementation, when obtaining the model loss based on the label data carried by each training sample and the obtained multiple prediction results corresponding to each training sample, the text analysis unit is specifically configured to:

[0037] Based on the label data carried by each of the training samples, combined with the obtained multiple prediction results corresponding to each of the training samples, obtain the sub-loss of each of the multiple recognition sub-models;

[0038] The model loss is obtained based on the sub-losses of the multiple recognition sub-models and the loss weights of the multiple recognition sub-models.

[0039] In one possible implementation, after obtaining the model loss based on the respective sub-losses of the multiple recognition sub-models and the respective loss weights of the multiple recognition sub-models, the text analysis unit is further configured to:

[0040] Based on the label data carried by each of the training samples, combined with the obtained multiple prediction results corresponding to each of the training samples, a model evaluation is performed to obtain a model evaluation value of each of the multiple recognition sub-models;

[0041] Based on the obtained model evaluation value, when determining to adjust the weight of at least one recognition sub-model among the multiple recognition sub-models, based on the loss weight of each of the at least one recognition sub-model in the current iteration process, the loss weight of each of the at least one recognition sub-model in the next iteration process is obtained.

[0042] In one possible implementation, when adjusting the model parameters based on the model loss, the text analysis unit is specifically configured to:

[0043] If the set parameter adjustment conditions are met, adjusting the other model parameters in the initial recognition model except for the model parameters whose parameter states are frozen based on the model loss, and adjusting the parameter states of the model parameters in the frozen state to a non-frozen state;

[0044] If the set parameter adjustment condition is not met, the other model parameters in the initial recognition model except for the model parameters whose parameter state is the frozen state are adjusted based on the model loss.

[0045] In a possible implementation, when constructing the target recognition model based on the trained initial recognition model, the text analysis unit is specifically configured to:

[0046] Based on the set target level, extracting the recognition sub-model corresponding to the target level from the trained initial recognition model;

[0047] A target recognition model is constructed based on the model parameters of the recognition sub-model corresponding to the target level.

[0048] In one possible implementation, each analysis result includes the predicted probability of each abnormal category of the corresponding complaint text;

[0049] Then, based on the obtained at least one analysis result, the account to be detected is identified. When the account identification result of the account to be detected is obtained, the account analysis unit is specifically configured to:

[0050] Based on the value of the at least one predicted probability obtained, selecting a predicted probability that meets the set evaluation conditions from the at least one predicted probability as a target probability;

[0051] Performing a weighted summation on the target probability and the average probability sum of the at least one predicted probability to obtain a weighted sum, and normalizing the weighted sum based on a set normalization parameter to obtain an abnormal probability of the account to be detected, wherein the normalization parameter is determined based on the number of complaint texts;

[0052] Based on the abnormal probability, an account identification result of the account to be detected is obtained.

[0053] In one possible implementation, when obtaining at least one complaint text generated for the account to be detected within a set collection period, the text obtaining unit is specifically configured to:

[0054] When the account to be detected meets the account detection conditions, at least one complaint text generated for the account to be detected is obtained according to the set collection period, where each complaint text is generated after the corresponding complaining account transfers resources with the account to be detected;

[0055] After identifying the account to be detected based on the at least one analysis result obtained and obtaining the account identification result of the account to be detected, the account analysis unit is further configured to:

[0056] If the account identification result of the account to be detected indicates that the account to be detected has abnormal resource transfer behavior, an alarm is issued for the account to be detected.

[0057] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0058] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which includes a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to perform the steps of any of the above methods.

[0059] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program, and the computer program is stored in a computer-readable storage medium. The processor of an electronic device reads and executes the computer program from the computer-readable storage medium, so that the electronic device performs the steps of any of the above methods.

[0060] In an embodiment of the present application, first, the complaint text generated for the account to be detected is used to identify the account to be detected. Since the complaint text usually contains key clues that the account to be retrieved has committed abnormal behavior, the use of the complaint text can ensure the accuracy of identifying abnormal accounts.

[0061] Secondly, artificial intelligence technology is used to analyze the abnormal behavior of complaint texts. Since the target recognition model is obtained through iterative training using historical complaint texts, it can quickly and accurately identify the abnormal behavior involved in the complaint text, thereby improving the recognition efficiency and recognition effect of abnormal accounts.

[0062] In addition, the target industry category to which the account to be tested belongs is used to assist in the analysis of abnormal behavior in complaint texts. Since complaint texts from different industries may contain industry-specific terms and expressions, the target industry category can be used to better identify abnormal behaviors related to the target industry category, especially abnormal behaviors unique to certain industries, thereby enhancing the adaptability and sensitivity of handling complaints from different industries and improving recognition accuracy.

[0063] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purposes and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0065] Figure 1 A schematic diagram of an application scenario provided in an embodiment of the present application;

[0066] Figure 2 A logical diagram of a data enhancement process provided in an embodiment of the present application;

[0067] Figure 3 A logical diagram of a tag recognition process provided in an embodiment of the present application;

[0068] Figure 4 A schematic diagram of an initial recognition model provided in an embodiment of the present application;

[0069] Figure 5 A schematic diagram of another initial recognition model provided in an embodiment of the present application;

[0070] Figure 6 A flow chart of a model training method provided in an embodiment of the present application;

[0071] Figure 7 A logical diagram of a loss weight adjustment process provided in an embodiment of the present application;

[0072] Figure 8 A schematic diagram of a target recognition model provided in an embodiment of the present application;

[0073] Figure 9 A schematic diagram of the difference between an abnormal account identification method provided in an embodiment of the present application;

[0074] Figure 10 This is a logical diagram of the abnormal account identification process provided in the embodiments of the present application;

[0075] Figure 11 This is a schematic diagram of the structure of an abnormal account identification device provided in an embodiment of the present application;

[0076] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0077] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of the technical solutions of this application, but not all of them. Based on the embodiments described in this application document, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the technical solutions of this application.

[0078] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0079] It is understandable that when the various embodiments of this application are applied to specific products or technologies, relevant licenses or consents need to be obtained, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0080] Below, some concepts involved in this application are first explained.

[0081] Sentence-BERT (SBERT): SBERT is a natural language processing model based on Bidirectional Encoder Representations from Transformers (BERT), which is optimized for sentence or paragraph level representations.

[0082] SBERT model structure: SBERT is built on the basis of BERT. BERT itself is based on the Transformer architecture and uses the self-attention mechanism to process input text.

[0083] SBERT's sentence embedding strategy: In SBERT, the input sentence is embedded in the context of the BERT network. Unlike BERT, SBERT is designed to output the embedding of the entire sentence, not just the representation features at the word (token) level.

[0084] SBERT's pooling layer: After obtaining token representation features from BERT, SBERT introduces a pooling operation to aggregate these token representation features into a single fixed-size sentence embedding. Common pooling strategies include average pooling and maximum pooling.

[0085] Siamese and Triple Network Architectures: During training, SBERT typically uses a Siamese or Triple Network architecture. This architecture involves passing multiple sentences through the same BERT network and then computing sentence embeddings, so that the embeddings of semantically similar sentences are closer, while the embeddings of semantically dissimilar sentences are further apart.

[0086] Fine-tuning for specific tasks: SBERT can be fine-tuned for sentence pair tasks such as semantic textual similarity, paraphrase identification, or other natural language processing tasks that require sentence embeddings.

[0087] Knowledge distillation using regression methods: Knowledge distillation is a model compression technique whose core idea is to train a small model (the student model) to mimic the behavior of a large, high-performance model (the teacher model). The teacher model is typically a large, deep, pre-trained model, while the student model is smaller and more computationally efficient. Traditional knowledge distillation typically uses the probabilistic output (the softmax output) from classification tasks as the medium for knowledge transfer. Knowledge distillation using regression methods, on the other hand, treats these probabilistic outputs as continuous values ​​and performs distillation using regression learning methods.

[0088] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0089] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, pre-trained models, operating / interaction systems, and mechatronics. Pre-trained models, also known as large models or basic models, can be fine-tuned and widely applied to downstream tasks across various AI disciplines. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0090] Computer vision (CV) is the science of making machines "see." Specifically, it refers to machine vision, where cameras and computers replace the human eye in identifying, detecting, and measuring objects. Further image processing is performed to transform the images into images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, aiming to build artificial intelligence systems that can extract information from images or multidimensional data. Large model technology has brought significant changes to the development of computer vision. Pre-trained models in the field of vision, such as the Swin Transformer, ViT, V-MOE, and MAE, can be fine-tuned to quickly and widely adapt to specific downstream tasks. Computer vision technologies generally include image processing, image recognition, image semantic understanding, image retrieval, optical character recognition (OCR), video processing, video semantic understanding, video content / action recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, and other technologies. Common biometric recognition technologies include facial recognition and fingerprint recognition.

[0091] Key technologies in speech technology include automatic speech recognition (ASR), text-to-speech (TTS), and voiceprint recognition. Enabling computers to hear, see, speak, and feel is the future direction of human-computer interaction, with speech becoming one of the most promising methods of human-computer interaction. Large model technology is revolutionizing the development of speech technology. Pre-trained models such as WavLM and UniSpeech, which leverage the Transformer architecture, possess strong generalization and versatility, enabling them to effectively handle a wide range of speech processing tasks.

[0092] Natural language processing (NLP) is an important field in the fields of computer science and artificial intelligence. It studies various theories and methods that enable effective communication between humans and computers using natural language. Natural language processing involves natural language, the language people use in daily life, and is closely related to linguistics research; it also involves computer science and mathematics. Pre-training models, an important technology for model training in the field of artificial intelligence, are developed from large language models (LLMs) in the field of NLP. After fine-tuning, large language models can be widely used in downstream tasks. Natural language processing technologies generally include text processing, semantic understanding, machine translation, robot question answering, knowledge graphs, and other technologies.

[0093] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning. Pretrained models are the latest development in deep learning, integrating these techniques.

[0094] A pre-training model, also known as a cornerstone model or a large model, refers to a deep neural network (DNN) with large parameters. It is trained on massive amounts of unlabeled data. Leveraging the function approximation capabilities of large-parameter DNNs, the pretrained machine learning (PTM) extracts common features from the data. Through techniques such as fine tuning, efficient parameter fine tuning (PEFT), and prompt-tuning, it is then adapted for downstream tasks. Therefore, pre-trained models can achieve ideal results in few-shot or zero-shot scenarios. Based on the data modality processed, PTMs can be categorized into language models (ELMO, BERT, GPT), vision models (swin-transformer, ViT, V-MOE), speech models (VALL-E), and multimodal models (ViBERT, CLIP, Flamingo, Gato). Multimodal models are those that represent features from two or more data modalities. Pre-trained models are important tools for outputting artificial intelligence generated content (AIGC) and can also serve as a universal interface for connecting multiple task-specific models.

[0095] Model compression and quantization: This refers to using compression and quantization techniques to reduce model size and accelerate model inference, thereby lowering model storage and computational costs. Model compression typically includes pruning, low-rank decomposition, and knowledge distillation. Model quantization converts floating-point parameters in the model to fixed-point or integer parameters, thereby reducing model size and accelerating model inference.

[0096] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, drones, digital twins, virtual humans, robots, artificial intelligence generated content (AIGC), conversational interaction, smart medical care, smart customer service, game AI, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0097] The solution provided in the embodiment of the present application mainly involves the machine learning technology of artificial intelligence. Specifically, after obtaining one or more complaint texts generated for the account to be detected, a target recognition model is used to analyze the abnormal behavior of the complaint text based on the target industry category to which the account to be detected belongs, and obtain the analysis results of the complaint text. Then, based on the obtained analysis results, the account to be detected is identified to obtain the account identification results of the account to be detected. The target recognition model can be obtained by fine-tuning the pre-trained model. The training and application process of the target recognition model is specifically described below and will not be repeated here.

[0098] In related technologies, in order to control risks, simple transaction limit controls are usually performed based on the industry category of the account. For example, accounts that trade physical items such as furniture are given larger transaction limits, while accounts that trade virtual items such as game props are given smaller transaction limits.

[0099] However, the above-mentioned transaction limit control is only based on the industry category of the account. It has a certain effect on normal accounts that comply with the law, but it is difficult to effectively control risks for abnormal accounts. For example, abnormal accounts can avoid transaction limit control by registering industry categories with larger transaction limits.

[0100] In an embodiment of the present application, first, the complaint text generated for the account to be detected is used to identify the account to be detected. Since the complaint text usually contains key clues that the account to be retrieved has committed abnormal behavior, the use of the complaint text can ensure the accuracy of identifying abnormal accounts.

[0101] Secondly, artificial intelligence technology is used to analyze the abnormal behavior of complaint texts. Since the target recognition model is obtained through iterative training using historical complaint texts, it can quickly and accurately identify the abnormal behavior involved in the complaint text, thereby improving the recognition efficiency and recognition effect of abnormal accounts.

[0102] In addition, the target industry category to which the account to be tested belongs is used to assist in the analysis of abnormal behavior in complaint texts. Since complaint texts from different industries may contain industry-specific terms and expressions, the target industry category can be used to better identify abnormal behaviors related to the target industry category, especially abnormal behaviors unique to certain industries, thereby enhancing the adaptability and sensitivity of handling complaints from different industries and improving recognition accuracy.

[0103] The following briefly introduces the application scenarios to which the technical solutions of the embodiments of the present application can be applied. It should be noted that the application scenarios described below are only used to illustrate the embodiments of the present application and are not limiting. In the specific implementation process, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.

[0104] The solution provided in the embodiments of this application can be applied to various account searches, such as identifying abnormal account numbers on shopping platforms, payment platforms, social platforms, etc. This solution can be applied as a basic technology in various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc.

[0105] See Figure 1 , which is a schematic diagram of an application scenario provided in an embodiment of the present application. The application scenario includes a terminal device 110 and a server 120. The number of terminal devices 110 can be one or more. The number of servers 120 can also be one or more. This application does not specifically limit the number of terminal devices 110 and servers 120.

[0106] In the embodiments of the present application, the terminal device 110 may be, but is not limited to, a smartphone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smartwatch, an IoT device, an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, or the like. The terminal device 110 may include a client for interacting with a corresponding account. The client may be in the form of, but is not limited to, an application, a mini-program, a webpage, or the like.

[0107] Server 120 is the backend server corresponding to the client. Server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0108] The terminal device 110 and the server 120 may be connected directly or indirectly via wired or wireless communication, which is not limited in this application.

[0109] The account search method mentioned in the embodiments of the present application can be executed by a server or a terminal device, or can be executed jointly by a server and a terminal device, without limitation.

[0110] For example, the server obtains at least one complaint text generated for the account to be detected within a set collection period, and based on the target industry category to which the account to be detected belongs, uses a target recognition model to perform abnormal behavior analysis on each of the at least one complaint texts, obtains corresponding analysis results, and then identifies the account to be detected based on the at least one obtained analysis result, and obtains an account recognition result for the account to be detected. For another example, the terminal device obtains at least one complaint text generated for the account to be detected within a set collection period, and based on the target industry category to which the account to be detected belongs, uses a target recognition model to perform abnormal behavior analysis on each of the at least one complaint texts, obtains corresponding analysis results, and then identifies the account to be detected based on the at least one obtained analysis result, and obtains an account recognition result for the account to be detected.

[0111] In the embodiment of the present application, the model used to identify abnormal behavior in complaint texts is called a recognition model. The training process of the recognition model is a process of multiple iterative training cycles using training samples. It can mainly include data preparation, model design, iterative training, model evaluation and other parts, which will be introduced below.

[0112] (1) Data preparation

[0113] In one possible implementation, a sample data includes a complaint text and its corresponding label data, wherein the label data is used to characterize the abnormal behavior category to which the historical complaint text belongs. In the embodiment of the present application, an abnormal behavior category can be regarded as a category label, and the process of performing abnormal behavior analysis on the complaint text can also be understood as the process of determining the category label of the complaint text. It should be noted that, hereinafter, for the sake of distinction, the complaint text in the sample data is referred to as the historical complaint text.

[0114] In one possible implementation, sample data includes, in addition to the historical complaint text and its corresponding label data, industry category text that characterizes the industry of the account involved in the complaint. In other words, the sample data consists of a sentence pair consisting of the historical complaint text and the industry category text. The following explanation uses sentence pairs as input.

[0115] In one possible implementation, the label data corresponding to the complaint text can be represented by a single-level label or a multi-level label.

[0116] The multi-level tag is composed of a plurality of category tags with a hierarchical relationship, and the hierarchical relationship between the plurality of category tags includes but is not limited to two or more levels. Referring to Table 1, it is an example of a multi-level tag provided in an embodiment of the present application. Each row in Table 1 represents a multi-level tag, and each multi-level tag is composed of a first-level tag and a second-level tag, and the second-level tag is a sub-tag of the first-level tag. The first-level tags include category tags such as "gambling", "pornography", and "fraud". Taking "gambling", "pornography", and "fraud" as examples, the sub-tags of the first-level tag "gambling" include "online gambling" and "offline gambling", the sub-tags of the first-level tag "pornography" include "pornographic videos", and "offline pornography", and the sub-tags of the first-level tag "fraud" include: "financial fraud", "e-commerce fraud", "brushing fraud", etc. The multi-level tags include: "gambling-online gambling", "gambling-offline gambling", "pornography-pornographic videos", etc.

[0117] It should be noted that this article only uses two levels as an example for illustration. In actual application, three or more levels can also be designed.

[0118] Table 1

[0119] (Multi-level tags)

[0120] First level label Second level label gamble Online gambling gamble Offline gambling pornography Porn Videos pornography Offline pornography fraud Financial fraud fraud e-commerce fraud fraud Fake order fraud fraud Cash-out fraud fraud loan fraud fraud Pornography fraud fraud gambling fraud Illegal pyramid schemes E-commerce pyramid schemes Illegal pyramid schemes Offline pyramid schemes Illegal fundraising Fundraising fraud Illegal fundraising Virtual trading platform fundraising Illegal fundraising Financial Management Illegal fundraising Sales rebates Transaction disputes Commodity price disputes Transaction disputes Product quality disputes Transaction disputes Delivery time disputes Transaction disputes Charge disputes

[0121] A single-level tag consists of a single category label. For example, a single-level tag could be a first-level tag such as "gambling," "pornography," or "fraud," or a second-level tag such as "online gambling," "offline gambling," or "pornographic videos."

[0122] For example, for the historical complaint text "This company is a big liar. They promised me high returns, but now I can't get a penny back. It's completely robbing me! This is my hard-earned money, and I'm losing money!", if multi-level labels are used, the label data corresponding to the historical complaint text will be "illegal fundraising-financial management". If single-level labels are used, the label data corresponding to the historical complaint text will be "illegal fundraising" or "financial management".

[0123] Since text classification methods usually only process single-level labels, they often perform poorly when processing hierarchical and multi-label data, especially when the sample distribution is uneven, and the classification effect is difficult to guarantee. Compared with using single-level labels for text classification, retaining high-level labels and the label hierarchy can improve the classification effect to a certain extent. In addition, using multi-level labels can transfer knowledge between label levels. When samples of a certain subdivision category are sparse, the relevant text classification knowledge can still be learned through the upper-level labels. Based on this, in actual applications, multi-level labels are usually used as label data for sample data. The following will only use multi-level labels as an example for explanation.

[0124] In one implementation, each sample data set can be divided into training samples and test samples, thereby obtaining a training sample set and a test sample set. The training samples are used for model fitting, i.e., the sample data used for training. Their primary purpose is to allow the model to learn the characteristics and patterns of the data, thereby determining the model parameters. The test samples are used to evaluate the model's ultimate generalization ability, i.e., the model's performance on unseen data, and their purpose is to verify the model's reliability and accuracy.

[0125] The following describes the sample data preparation process and labeling process respectively.

[0126] (1) Sample data preparation process

[0127] In one embodiment, initial samples are constructed based on each historical complaint text and its corresponding industry type text. Each initial sample includes a historical complaint text and its corresponding industry type text.

[0128] Furthermore, the initial sample can be directly used as sample data, or data enhancement can be performed on the initial sample, and the initial sample and the expanded sample obtained by data enhancement can be used as sample data.

[0129] Specifically, during data augmentation, each initial sample can be augmented according to a predefined data augmentation method to obtain each expanded sample. Sample data can then be obtained based on each initial and expanded sample. See below for details on data augmentation methods. Data augmentation can prevent the model from learning information irrelevant to the target, thereby reducing the risk of over-optimization within the training sample set and improving the model's robustness and generalization performance.

[0130] In one implementation, the historical complaint texts may be complaint texts for known abnormal accounts. Abnormal accounts include, but are not limited to, accounts that engage in abnormal activities such as illegal fundraising, fraud, and gambling. For example, in the daily abnormality identification process, the detected abnormal accounts are marked, for example, with multi-level labels such as "gambling - online gambling", "gambling - offline gambling", and "pornography - pornographic videos". Then, the complaint texts generated for each abnormal account are extracted, and based on the extracted complaint texts, the industry type texts representing the industry type to which the abnormal account belongs are combined to form each initial sample, and the multi-level labels marked for the abnormal account are used as the multi-level labels of the initial sample formed according to the complaint texts of the corresponding abnormal account.

[0131] In one implementation, negative samples can be constructed using complaint texts from known trusted accounts. Trusted accounts are those with non-abnormal behaviors such as service quality issues and logistics problems. For example, from complaint texts from trusted accounts involving non-abnormal behaviors such as service quality issues and logistics problems, some or all complaint samples are selected randomly or according to a predefined screening method (for example, selecting historical complaint samples within a certain period). These selected complaint samples, combined with the corresponding industry type text and category labels, are used as negative samples.

[0132] Refer to Table 2, which is an example of the initial sample provided in the embodiments of the present application. In the initial sample 1, the historical complaint text is "This company is a big liar. The high returns promised are not available now. It's completely robbing money! This is my hard-earned money, and I am losing money!", the industry type text is "financial management", and the multi-level label is "illegal fundraising-financial management"; in the initial sample 2, the historical complaint text is: "I was completely cheated by this mini program. I bought a watch, paid for it, but it was not shipped! I went to them to argue, and the customer service seemed to have disappeared. I think this is a scam. Don't let anyone be fooled again!", the industry type text is "e-commerce", and the multi-level label is "fraud-e-commerce fraud"; in the initial sample 3, the historical complaint text is "inducing users to recharge, it's just a manipulated gambling game that you can't win at all", the industry type text is "online games", and the multi-level label is "gambling-online gambling".

[0133] Table 2

[0134] (Initial sample example)

[0135]

[0136] Below, the data enhancement process is described using the initial sample x as an example. The initial sample x can be any initial sample among the initial samples. Specifically, data enhancement can be performed using at least one of the following operations, but not limited to:

[0137] Data enhancement operation 1: synonym replacement.

[0138] Specifically, at least one word in the historical complaint text in the initial sample x is replaced with at least one synonym of each word to obtain an expanded sample of the initial sample x. Synonyms can be words with the same or similar semantics.

[0139] In this embodiment of the present application, the words selected from the historical complaint text x include, but are not limited to, nouns, verbs, adjectives, and the like. The number of words selected is not limited. For example, if the historical complaint text contains the word "gambling," the word "gambling" can be replaced with its synonym "gambling money" to obtain an expanded sample.

[0140] Generally speaking, in order to keep the semantics of the sentence unchanged, it is possible to avoid replacing key words in the sentence, which include but are not limited to account names, etc. In the embodiment of the present application, the method for determining key words is not limited and will not be described in detail here.

[0141] In one implementation, a dictionary resource such as WordNet or Chinese Synonym Cilin can be used to obtain synonyms of a word. For example, using Chinese Synonym Cilin, synonyms of “scam” can be obtained as “cheat” or “fraud”.

[0142] When performing replacement, if a word corresponds to a synonym, then the word can be directly replaced with the synonym of the word. If a word corresponds to multiple synonyms, then a synonym can be randomly selected from the multiple synonyms corresponding to the word to replace the word. Alternatively, based on the frequency of word usage, a synonym with a frequency exceeding a set threshold can be selected from the multiple synonyms corresponding to the word to replace the word. Of course, other methods can also be used to select synonyms, and this is not limited to this. In addition, if a word corresponds to multiple synonyms, multiple synonyms can also be used to replace the word, thereby obtaining multiple expanded samples.

[0143] In one implementation, when making replacements, attention should also be paid to the contextual adaptability of words to ensure that the replaced words are semantically reasonable in the sentence. For example, after generating an expanded sample, the expanded sample can also be semantically verified, and when the expanded sample passes the semantic verification, the expanded sample can be used as the initial sample.

[0144] Taking the initial sample 3 as an example, in the initial sample 3, the historical complaint text is "Inducing users to recharge. It's just a rigged money-making game that can't be won at all", the industry type text is "Online games", and the multi-level label is "Gambling - Online gambling". For the word "gambling" contained in the historical complaint text, replace it with its synonym "cheating money", and obtain the replaced complaint text "Inducing users to recharge. It's just a rigged cheating money game that can't be won at all". The multi-level label of the augmented sample is "Gambling - Online gambling". Then, based on the replaced complaint text, combined with the industry type text and multi-level label of the initial sample 3, obtain the augmented sample 1 of the initial sample 3. The historical complaint text in the augmented sample 1 is "Inducing users to recharge. It's just a rigged cheating money game that can't be won at all", the industry type text is "Industry: Online games", and the multi-level label is "Gambling - Online gambling".

[0145] Data augmentation operation 2: Interference word replacement.

[0146] Specifically, replace the keywords contained in the historical complaint text of the initial sample x with interference words that are similar in font to the keywords, and obtain the augmented sample of the initial sample x.

[0147] Among them, the interference words that are similar in font to the keywords include but are not limited to: misspelled words of the keywords, pinyin, traditional Chinese characters, etc. The number of keywords can be one or more. For example, assume that the historical complaint text contains the word "gambling", and replace it with "block money" or "du qian".

[0148] Still taking the initial sample 3 as an example, in the initial sample 3, the historical complaint text is "Inducing users to recharge. It's just a rigged money-making game that can't be won at all", the industry type text is "Online games", and the multi-level label is "Gambling - Online gambling". For the word "gambling" contained in the historical complaint text, replace it with "block money", and obtain the replaced complaint text "Inducing users to recharge. It's just a rigged block money game that can't be won at all". The label corresponding to the augmented sample is "Gambling - Online gambling". Then, based on the replaced complaint text, combined with the industry type text and label of the initial sample 3, obtain the augmented sample 2 of the initial sample 3. The historical complaint text in the augmented sample 2 is "Inducing users to recharge. It's just a rigged block money game that can't be won at all", the industry type text is "Online games", and the multi-level label is "Gambling - Online gambling".

[0149] In data augmentation operation 2, by replacing some keyword vocabularies with their common misspelled words or traditional Chinese characters, etc., it helps the model learn more robust features.

[0150] Data augmentation operation 3: Semantic augmentation.

[0151] Specifically, based on the semantic information of the historical complaint text in the initial sample x, the initial sample x is expanded with similar texts to obtain an expanded sample of the initial sample x.

[0152] In one implementation, a large language model (LLM) is used to expand the historical complaint text in the initial sample x with similar text, thereby obtaining an expanded sample of the initial sample x. For example, the initial sample x is used as a seed and fed into the LLM. LLM's powerful text understanding and generation capabilities are then used to generate complaint texts that are semantically similar to the historical complaint texts. Based on the generated complaint texts, expanded samples of the initial sample x are then obtained.

[0153] It should be noted that the label data of the expanded sample of the initial sample x generated by data augmentation operations 1-3 is consistent with the label data of the initial sample x. The industry type text of the expanded sample of the initial sample x generated by data augmentation operations 1-3 is consistent with the industry type text of the initial sample x.

[0154] In one implementation, semantic and grammatical checking can also be performed on historical complaint texts to ensure that the intent and emotional tone of the complaint text in the replacement sample are consistent with those in the initial sample. Semantic and grammatical checking includes, but is not limited to, one or more of semantic consistency and grammatical correctness. Both semantic consistency and grammatical correctness can be performed manually or through automated scripts, but are not limited to these methods.

[0155] In an embodiment of the present application, data enhancement of initial data can be achieved by using one or more of the three data enhancement operations mentioned above.

[0156] The proportion of augmented samples generated by data augmentation in the total training samples or total augmented samples can be adjusted based on actual conditions. If two or more data augmentation operations are used, the number of augmented samples generated by each data augmentation operation can be the same or different, and there is no restriction on this.

[0157] As an example, the augmented samples obtained by different data augmentation operations are used as different types of augmented samples. The sample ratio of each type of augmented sample in the total augmented samples is set. Then, based on the sample ratio of each type of augmented sample, data augmentation is performed on the initial sample to obtain each augmented sample.

[0158] For example, see Figure 2As shown, assuming that the sample ratio of the expanded samples generated by data augmentation operation 1 (synonym replacement), data augmentation operation 2 (interference word replacement), and data augmentation operation 3 (semantic supplement) is 6:4:4, for each initial sample, 60 expanded samples can be generated by data augmentation operation 1, 40 expanded samples can be generated by data augmentation operation 2, and 40 expanded samples can be generated by data augmentation operation 3 according to 6:4:4. Taking the first three initial samples in each initial sample as an example, the historical complaint text in the first initial sample is replaced with synonyms, interference words and semantic supplemented to obtain the corresponding historical complaint text. Then, the multi-level labels and industry type text in the first initial sample are used to combine the three obtained historical complaint texts to construct three supplementary samples; the historical complaint text in the second initial sample is replaced with synonyms to obtain the corresponding historical complaint text. Then, the multi-level labels and industry type text in the first initial sample are used to combine the obtained historical complaint text to construct a supplementary sample; the historical complaint text in the third initial sample is replaced with synonyms and interference words to obtain the corresponding historical complaint text. Then, the multi-level labels and industry type text in the third initial sample are used to combine the two obtained historical complaint texts to construct two supplementary samples.

[0159] (2) Labeling process of sample data

[0160] In real-world business scenarios, it's easy to accumulate massive amounts of unlabeled complaint text. Effectively leveraging this text is crucial for identifying anomalous accounts. However, fine-tuning pre-trained models requires a large amount of labeled text, which is time-consuming, labor-intensive, and costly.

[0161] Based on this, in an embodiment of the present application, a weak supervision method that combines manual few-shot learning and self-consistency prompt is used to distill the knowledge of the large language model (LLM) into the feature extraction network of the recognition model (such as Sentence-BERT), that is, the label data output by the LLM is used, and the label data is spliced ​​with the historical complaint text as supervised sample data for fine-tuning the recognition model, thereby improving the model effect when the labeled data is limited. Among them, the soft label refers to the probability of belonging to a certain category represented by a probability distribution. For example, for a three-category task, the soft label of a sample data is [0.3, 0.5, 0.2], indicating that the probability of the sample data belonging to the three categories is 30%, 50%, and 20% respectively.

[0162] In some implementations, the LLM model involves a model training phase and a model application phase. During the model training phase, the LLM is iteratively trained using labeled sample data until the model converges, obtaining a trained LLM. Due to the limited amount of labeled text, the LLM model is trained using few-shot learning, which can reduce the resources consumed by labeling while ensuring the LLM model effect. During the model application phase, the trained LLM is used to label unlabeled sample data to obtain a labeled sample dataset. In this article, the LLM can also be referred to as a label annotation model.

[0163] In some implementations, whether it is the LLM training process or the LLM application process, the label data output by the LLM can be represented in the form of soft labels. A soft label contains the label probability that the corresponding historical complaint text belongs to each category label, that is, a soft label contains the label probability that the corresponding training sample belongs to each category label. Of course, in actual application, the label data can also be represented in the form of hard labels.

[0164] In one possible implementation, taking the LLM model application process as an example, during the labeling process, the label data of a sample data is generated in the following way:

[0165] Using the LLM model, a training sample is labeled multiple times to obtain the corresponding labeling results of the multiple labelings. Each labeling result represents the abnormal behavior category to which the training sample belongs.

[0166] Based on the obtained multiple annotation results, the results are aggregated to obtain a soft label of a training sample, and the soft label is used as the label data of a training sample, wherein the soft label includes the label probability that a training sample belongs to each abnormal behavior category.

[0167] Among them, when using the LLM model to perform multiple label annotations on a training sample, as a possible implementation method, the historical complaint text in the training sample can be input into the LLM model to obtain the annotation results corresponding to the input historical complaint text, and the annotation results corresponding to the historical complaint text can be used as the annotation results of the sample data.

[0168] As another possible implementation method, the historical complaint text and industry type text in the training sample can also be input into the LLM model to obtain the annotation results corresponding to the input historical complaint text and industry type text, and the annotation results corresponding to the historical complaint text and industry type text can be used as the annotation results of the sample data.

[0169] Whether it is the LLM model training process or the LLM model application process, when aggregating results based on multiple obtained annotation results, for each abnormal behavior category in each abnormal behavior category, the proportion of annotation results that represent the training samples belonging to the abnormal behavior category in the multiple obtained annotation results is used as the label probability of the abnormal behavior category.

[0170] For example, see Figure 3 As shown in the figure, for a historical complaint text "This platform allows you to charge money to play, which looks like gambling.", the label recognition model is used to make 5 inferences. Assuming that the abnormal behavior categories include: "gambling", "fraud", and "illegal fundraising", the label prediction results of the 5 inferences are: "gambling" 3 times, "fraud" 1 time, and "illegal fundraising" 1 time. Then, the soft labels of the historical complaint text are [3 / 5, 1 / 5, 1 / 5], where 3 / 5 represents the label probability corresponding to gambling, and the two 1 / 5s represent the label probabilities of illegal fundraising and fraud respectively.

[0171] Specifically, knowledge distillation based on regression methods includes but is not limited to the following steps:

[0172] Step A: Construct prompts. Design prompts to guide the LLM in generating categorization labels for complaint texts. For example, a prompt might read, "The following is a complaint text. Please determine its category based on the content: fraud, illegal fundraising, gambling, etc."

[0173] Step B: Few-shot learning: Using a small number of examples, we help the LLM understand how to classify complaints based on their content.

[0174] Examples consist of historical complaint text and its corresponding multi-level labels. For example, a complaint like "I bought a watch on this website, but received a fake. It's a complete scam!" would have a multi-level label of "fraud." Another example would be a complaint like "This company promised high returns on investment, but I suspect it's illegal fundraising." The multi-level label would be "illegal fundraising."

[0175] In the embodiment of the present application, due to the limited amount of annotated text, the LLM model is trained using few-shot learning. This involves training the LLM using a small amount of annotated text, and using the trained LLM to annotate unannotated sample data. This allows the full utilization of the vast amount of unannotated complaint text to train the recognition model, thereby improving the model's recognition performance. In actual applications, if the amount of annotated text is not limited, other training methods may also be used, and this is not a restriction.

[0176] Moreover, by using knowledge distillation of regression methods, the knowledge of LLM is effectively transferred to the recognition model. The recognition model can learn and imitate the prediction behavior of LLM. In this way, some text comprehension capabilities of LLM are retained while reducing resource consumption, making it possible to deploy high-performance models with limited hardware resources.

[0177] In addition, when using LLM to predict the label data of historical complaint texts, multiple inferences are used to obtain the soft label distribution of each sample. The soft label distribution obtained by multiple inferences can provide richer training signals for the recognition model, thereby improving the performance and generalization ability of the recognition model on specific tasks.

[0178] (2) Model design

[0179] In one implementation, the recognition model consists of a feature extraction network and a classification network. The feature extraction network extracts text features from the complaint text and industry category text, while the classification network generates corresponding prediction results based on the input text features. For ease of distinction, the recognition model used in the training phase is referred to as the initial recognition model, and the applied recognition model is referred to as the target recognition model.

[0180] In one implementation, Sentence-BERT can be used as the base model in the recognition model to extract text features, that is, Sentence-BERT is used as the feature extraction network. In actual applications, pre-trained models using other architectures can also be used, such as those based on a generative pre-trained transformer (GPT) or a language model (XLNet). In addition, pre-trained models with multiple different architectures can be combined to obtain richer text feature representations through model fusion techniques. Alternatively, models pre-trained for specific fields (such as finance and law) can be used to provide more accurate text understanding capabilities within that field.

[0181] In one implementation, a multi-objective learning task is designed for model training under a multi-level label structure, and the classification of each level of labels is used as a separate learning objective. For example, the classification of the first level of labels is used as learning objective 1, and the classification of the second level of labels is used as learning objective 2. For each learning objective in the multi-objective learning task, an independent Sentence-BERT is used as an encoder to extract text features.

[0182] In one implementation, for each learning objective, a corresponding multilayer perceptron (MLP) is configured as a classification network. For example, a two-layer MLP is configured as the classification network. The classification network receives text features from the encoder as input and outputs a prediction result for each category label. Exemplarily, the prediction result can be the predicted probability of the corresponding category label. In the embodiment of the present application, a regression-based classification method is used, and therefore, an activation function (softmax) is not used to transform the model output.

[0183] For each label level, define the output prediction score as y i ,y i It can be expressed by formula (1):

[0184] y i =MLP(SBERT(x i )) Formula (1)

[0185] Among them, x_i is the input text, which includes industry type text and historical complaint text, SBERT is the Sentence-BERT encoder, and MLP is the multi-layer perception layer.

[0186] See Figure 4 As shown, it is a structural diagram of an initial recognition model provided in an embodiment of the present application. The initial recognition model is composed of two recognition sub-models (represented by dotted boxes). Recognition sub-model 1 is used to recognize first-level labels, and recognition sub-model 2 is used to recognize second-level labels. Recognition sub-model 1 and recognition sub-model 2 are both composed of a feature extraction network and a classification network. In this article, the feature extraction network can also be called a feature extraction layer, and the classification network can also be called a classification layer.

[0187] It should be noted that in actual application, three or more recognition sub-models can be deployed in the initial recognition model according to the number of learning tasks. In the embodiment of the present application, only two recognition sub-models are used as an example for illustration, and there is no limitation on this.

[0188] In some implementations, the feature extraction network (i.e., the feature extraction layer) uses a dual Sentence-BERT structure. Figure 5 As shown in the figure, the Sentence-BERT structure on the left is used to process the first-level labels, and the output dimension D1 of its linear layer matches the number of first-level labels. The Sentence-BERT structure on the right is used to process the second-level labels, and the output dimension D2 of its linear layer matches the number of second-level labels. <D2。

[0189] The classification network consists of an input layer and an output layer. The input layer accepts the sentence embedding vectors (i.e., text features) extracted by SBERT. These vectors are fixed-length floating-point arrays that represent the features of the input sentences. The output layer should match the dimensions of the soft labels. It should be noted that the dimensions corresponding to each level label in the multi-level label are usually different, and the corresponding output layer dimensions are selected for labels at different levels. In one implementation, one or more fully connected hidden layers can be added between the input layer and the output layer of the classification network to improve the learning ability of the model. Each hidden layer should use an activation function (such as ReLU) to increase nonlinearity.

[0190] In one implementation, a parameter sharing strategy can be used to reduce learning difficulty and the number of model parameters. Taking the dual Sentence-BERT structure as an example, the parameter sharing strategy can adopt, but is not limited to, any of the following strategies:

[0191] Parameter Sharing Strategy 1: Use Sentence-BERT's Embedding layer as a shared layer. Use Sentence-BERT as the feature extraction basis for all classification tasks. This means all tasks will use the same word embedding weights, but different attention layer weights.

[0192] Then, we design specific classification layers for each level of the task. Because each level of classification may have its own specific requirements and characteristics, we design customized intermediate and output layers for each task on top of the shared base layer. For example, the first level of labels may require a simple classification layer, while the second level of labels may require a more complex network structure to capture more nuanced differences.

[0193] Parameter Sharing Strategy 2: Independent Parameter Strategy. Each classification task uses an independent model architecture without sharing underlying parameters. This allows for specialized optimization of each task, potentially improving performance on that specific task. Furthermore, ensemble learning can be used to combine the model results from different tasks to improve overall classification performance.

[0194] Parameter Sharing Strategy 3: Hybrid Parameter Sharing. This strategy partially shares model parameters while retaining some independent parameters for each task. This approach balances the sharing and independence of model parameters and allows the model structure to be adjusted based on task characteristics.

[0195] See Figure 5 As shown in FIG, which is a schematic diagram of the structure of an initial recognition model provided in an embodiment of the present application. The initial recognition model adopts a dual Sentence-BERT structure. The Sentence-BERT structure on the left is used to process the first-level labels, and the Sentence-BERT structure on the right is used to process the second-level labels.

[0196] In each Sentence-BERT architecture, the Input Embedding layer converts the input text sequence into embeddings. Positional encodings are then added to the embeddings. These positional encodings are then fed into a multi-head self-attention layer, repeated 12 times, to obtain text features. The multi-head self-attention layer consists of multi-head attention, residual and normalization (add&norm), a feedforward network, and residual and normalization (add&norm).

[0197] The text features output by the Sentence-BERT structure on the left are input into the classification network on the left (i.e., two-layer MLP) to obtain the prediction results of the first-level labels.

[0198] The text features output by the Sentence-BERT structure on the right are input into the classification network on the right (i.e., two-layer MLP) to obtain the prediction results of the first-level labels.

[0199] (3) Iterative training

[0200] In the embodiment of the present application, after the training data is prepared, the constructed model can be trained using the training data.

[0201] In one embodiment, the parameters and data required for the initial recognition model, including the parameters to be trained, can be set based on the structure of the aforementioned model. After setting hyperparameters such as batch, number of epochs, and learning rate, training can begin, ultimately yielding a trained initial recognition model. For example, setting the batch size of the initial recognition model to 128, the epoch to 1000, and the learning rate to 0.0001 means iterating training 1000 times, with each iteration dividing the training samples into 128 batches for learning.

[0202] See also Figure 6 , which is a flow chart of the initial recognition model training method provided in an embodiment of the present application.

[0203] S601: Construct a training sample set based on each historical complaint text. Each training sample carries corresponding label data. The label data is used to characterize the multiple abnormal behavior categories to which the training sample belongs. There is a hierarchical relationship between the multiple abnormal behavior categories. The details of S601 are described in the data preparation stage and will not be repeated here.

[0204] S602: Based on the training sample set, iteratively train the initial recognition model to obtain a trained initial recognition model.

[0205] During the iterative training process, all training samples are divided into specified batches, and training is performed based on the training samples of each batch. Since the steps performed for each batch in each iteration are similar, the training for one batch is used as an example here.

[0206] S6021. Obtain a batch of training samples, and use multiple recognition sub-models to perform label prediction on each training sample to obtain the corresponding prediction results of each training sample at multiple levels.

[0207] The prediction result corresponding to a training sample at a level is used to represent the level label of the level to which the training sample belongs. In some embodiments, if the labeled data is represented by soft labels, the prediction result should match the dimension of the soft labels.

[0208] Specifically, when using multiple recognition sub-models to predict labels for each training sample, the following operations are performed for each of the multiple recognition sub-models:

[0209] Using a feature extraction network in a recognition sub-model, feature extraction is performed on each training sample to obtain corresponding text features;

[0210] The classification network in the recognition sub-model is used to classify the obtained text features respectively, and obtain the prediction results corresponding to each training sample at the corresponding level.

[0211] In an embodiment of the present application, if the label data of the sample data is represented by a soft label, then the prediction result includes the prediction score of each category label.

[0212] See Figure 4The model architecture of the initial recognition model shown in the figure is: the feature extraction network and classification network on the left side of the initial recognition model are recognition sub-model 1, and the feature extraction network and classification network on the right side are recognition sub-model 2. Recognition sub-model 1 is used to identify first-level labels, and recognition sub-model 2 is used to identify second-level labels. In one iteration, the feature extraction network in recognition sub-model 1 is used to extract features from each training sample, obtaining the text features corresponding to each training sample. These features are then input into recognition sub-model 1 to obtain prediction scores for first-level labels, such as "gambling," "pornography," and "fraud." Furthermore, the feature extraction network in recognition sub-model 2 is used to extract features from each training sample, obtaining the text features corresponding to each training sample. These features are then input into recognition sub-model 2 to obtain prediction scores for second-level labels, such as "online gambling," "offline gambling," and "pornographic videos."

[0213] In the above implementation, the classification of each level of labels is taken as a separate learning objective, and each objective uses an independent feature network and classification layer for text feature extraction and category recognition. Through multi-objective learning, the knowledge learned in one level can be transferred to other levels, thereby improving the performance and generalization ability of the recognition model.

[0214] S6022: Based on the label data carried by each training sample and in combination with the multiple prediction results corresponding to each training sample, obtain the model loss.

[0215] Specifically, when executing S6022, perform the following operations:

[0216] Based on the label data carried by each training sample, combined with the multiple prediction results corresponding to each training sample, the sub-loss of each of the multiple recognition sub-models is obtained;

[0217] Based on the sub-losses of multiple recognition sub-models and the loss weights of multiple recognition sub-models, the model loss is obtained.

[0218] In the embodiment of the present application, the loss weight can be a dynamic parameter or a fixed parameter.

[0219] In one implementation, the loss weight is treated as a dynamic parameter that is adjusted during training. Choosing a dynamic weighting strategy to adjust the weights of the loss functions at different levels is a reasonable choice, especially for hierarchical multi-label text classification tasks, where different levels may have different levels of difficulty and importance.

[0220] Considering that dynamically adjusting the weights of losses at different levels may cause the model to encounter convergence problems during training, especially if the weight adjustment strategy is inappropriate or too aggressive. The reasons why this may happen include but are not limited to: over-adjusting weights. If the weight of a certain level is excessively increased, it may cause the model to pay too much attention to that level, thereby ignoring the importance of other levels, resulting in a decline in overall performance; rapidly changing gradients. Frequent or drastic adjustments to loss weights may cause the gradient of the loss function to change rapidly, making the optimization process difficult and difficult to find a stable descent path; unstable training. Rapid changes in weights may cause fluctuations during training, making it difficult for the model to stably learn effective features. In order to avoid the above situation, in an embodiment of the present application, a moving average smoothing weight adjustment method is used to adjust the loss weight:

[0221] Specifically, based on the label data carried by each training sample, a model evaluation is performed in combination with the multiple prediction results corresponding to each training sample to obtain the model evaluation values ​​of each of the multiple recognition sub-models;

[0222] Based on the obtained model evaluation value, when determining to adjust the weight of at least one recognition sub-model among multiple recognition sub-models, based on the loss weight of at least one recognition sub-model in the current iteration process, the loss weight of at least one recognition sub-model in the next iteration process is obtained.

[0223] Among them, the obtained model evaluation value determines the weight adjustment of at least one recognition sub-model among multiple recognition sub-models, including but not limited to: if the model evaluation value of the recognition sub-model represents that the performance improvement of the recognition sub-model is slow (for example, the increase in the performance evaluation value within the set time period is lower than the set threshold), then the loss weight of the recognition sub-model is increased, and then based on the loss weight of the recognition sub-model in the current iteration process, the loss weight of the recognition sub-model is increased.

[0224] See Figure 7 As shown, the specific implementation of the weight adjustment method based on moving average smoothing is as follows:

[0225] First, initialize the weights. Use L i Represents the loss weight corresponding to the learning task of the i-th level, that is, the loss of each identification sub-model, which is L i Assigning initial loss weights The initial loss weight is set to where α i is a hyperparameter based on the difficulty and importance of the learning task at the i-th level.

[0226] Then, at the end of each training cycle (epoch), the performance of the model in each multi-objective learning task is evaluated (Pi ), that is, the performance of each recognition sub-model. i Evaluation indicators such as accuracy or F1 score can be used but are not limited to them.

[0227] Then, weight adjustment and smoothing are performed. Specifically, the weight of each level loss is dynamically adjusted according to the performance evaluation results. That is, based on the model evaluation values ​​of multiple recognition sub-models, the loss weight of the recognition sub-model is adjusted according to the set adjustment strategy. The adjustment strategy is: if L i If the performance of i , then, use the exponential moving average (EMA) to smooth the increased loss weight and obtain the next iteration process, L i The loss weight of the next iteration, L i The loss weight can be expressed by formula (2):

[0228]

[0229] Among them, γ is the smoothing coefficient, and the right side of the equation is the adjusted but unsmoothed loss weight at time t, and the left side of the equation is the adjusted and smoothed weight at time t, is the adjusted and smoothed weight at time t-1, that is, the loss weight based on the current iteration process.

[0230] Furthermore, in the loss calculation of each training batch, the total loss L is calculated based on the loss weights of the current multiple recognition sub-models. total , where L total =∑ i w i ×L i , and then gradient descent optimization is performed using the updated and smoothed loss weights.

[0231] In the above implementation, during the training process, the focus on tasks at different levels can be flexibly adjusted, thereby improving the overall performance of the model in classification tasks with multi-level labels.

[0232] In some implementations, fixed weight assignments can be used, where the loss weights are fixed. By pre-setting and fixing the loss weights for multiple objective tasks, this can simplify the training process and provide a stable learning environment in some cases. In practical applications, the uncertainty of the model's predictions on different tasks can be considered, and greater weights can be assigned to tasks with higher uncertainty to promote model learning on these tasks.

[0233] In one possible implementation, for regression tasks, the mean square error (MSE) loss function is often used for loss calculation. MSE calculates the squared difference between the model's predicted value and the actual label and is applicable to continuous outputs. MSE can be calculated using formula (3):

[0234] MSE=(1 / n)*Σ(actual-predicted) 2 Formula (3)

[0235] Among them, Σ represents summation, n represents the number of sample data, actua represents the label data of the sample data, and predicted represents the predicted result of the sample data.

[0236] Of course, in actual applications, other loss functions can also be used for loss calculation, such as cross entropy loss function, quadratic loss function, and absolute loss function, without restriction.

[0237] S6023: Determine whether the model convergence condition is met. If so, execute S6025; otherwise, execute S6024.

[0238] In the embodiment of the present application, the convergence condition may include at least one of the following conditions:

[0239] (1) The total loss value is not greater than the preset loss value threshold.

[0240] (2) The number of iterations reaches the preset upper limit.

[0241] If the above conditions are met, it is determined that the initial recognition model has met the convergence conditions, and the training ends. Otherwise, it is determined that the initial recognition model has not met the convergence conditions, then it is necessary to continue to adjust the model parameters and use the adjusted initial recognition model to enter the next training process.

[0242] S6024. Adjust model parameters based on model loss.

[0243] In some implementations, backpropagation is used to optimize model parameters. Common optimizers such as Adam or AdamW can be used.

[0244] Specifically, when executing S6024, the following methods may be used, but are not limited to:

[0245] If the set parameter adjustment conditions are met, then based on the model loss, the model parameters in the initial recognition model except for the model parameters whose parameter states are frozen are adjusted, and the parameter states of the model parameters in the frozen state are adjusted to the non-frozen state;

[0246] If the set parameter adjustment conditions are not met, the other model parameters in the initial recognition model except for the model parameters whose parameter states are frozen are adjusted based on the model loss.

[0247] The parameter adjustment conditions include, but are not limited to, the current number of iterations reaching the set unblocking threshold or the completion of an epoch of training. When the parameter state of a model parameter is frozen, gradients are not calculated for the model parameter and no updates are performed. When the parameter state of a model parameter is not frozen, gradients are calculated for the model parameter and updates are performed.

[0248] For example, before model training, the parameter states of some shared layer parameters (such as the word embedding layer) are frozen, thereby preserving the semantic feature extraction capabilities of the pre-trained model in the early stages of model training. As training progresses, for example, after training an epoch, the parameter states of the frozen model parameters are adjusted to a non-frozen state to unfreeze the model parameters. This prevents drastic changes in lower-layer parameters caused by initial fine-tuning.

[0249] In the above implementation, by freezing some model parameters, these frozen model parameters do not need to be updated during the training process, thereby reducing training time and accelerating the model training process. In addition, when using a pre-trained model as the initial recognition model, freezing some of the pre-trained model parameters can maintain the feature extraction capabilities of these parameters.

[0250] S6025. Output the initial recognition model.

[0251] S603: Build a target recognition model based on the trained initial recognition model.

[0252] In some implementations, after the initial recognition model is trained, the initial recognition model can be directly used as the target recognition model. However, in actual application, in order to save the resources required to deploy the model, in an embodiment of the present application, some sub-models in the initial recognition model can be retained as the target recognition model.

[0253] Specifically, based on the trained initial recognition model, a target recognition model is constructed, including:

[0254] Based on the set target level, the recognition sub-model corresponding to the target level is extracted from the trained initial recognition model; and based on the model parameters of the recognition sub-model corresponding to the target level, the target recognition model is constructed.

[0255] For example, see Figure 8 As shown in the figure, if the target level is the first level, then based on the set target level, recognition sub-model 1 is extracted from the trained initial recognition model. Recognition sub-model 1 is used to recognize the first-level labels, and then the target recognition model is constructed based on the model parameters of recognition sub-model 1. If the target level is the second level, then based on the set target level, recognition sub-model 2 is extracted from the trained initial recognition model. Recognition sub-model 2 is used to recognize the first-level labels, and then the target recognition model is constructed based on the model parameters of recognition sub-model 2.

[0256] In the above implementation, by deploying some sub-models in the trained initial recognition model, the resources required for deploying the model, such as storage resources, computing resources, etc., can be greatly reduced. In addition, by retaining sub-models for the target level, the model application effect can be guaranteed when resources are limited.

[0257] (IV) Model Evaluation

[0258] In multi-label classification problems, due to class imbalance, a single accuracy metric may not fully reflect the model's performance. In this case, using multiple metrics for evaluation can provide a more comprehensive overview of performance. Evaluation metrics suitable for multi-label text classification problems include, but are not limited to, F1 score, macro-average F1 score, and micro-average F1 score.

[0259] For example, the F1 score can be expressed using formula (4):

[0260]

[0261] Among them, P represents the precision. R represents the recall rate (Recall), TF stands for true positive, FP stands for false positive, and FN stands for false negative.

[0262] For example, the macro-average F1 score can be expressed using formula (5):

[0263]

[0264] Among them, F1 macro represents the macro average F1 score, N represents the total number of abnormal behavior categories, and F1 iis the F1 score of the i-th abnormal behavior category.

[0265] For example, the micro-average F1 score can be expressed using formula (6):

[0266]

[0267]

[0268]

[0269] Among them, F1 micro represents the micro-average F1 score, P micro Using formula (7), R micro It is expressed as formula (8). In formula (7) and formula (8), TP total =∑TP i , FP total =∑FP i , FN total =∑FN i , TF represents true positives, FP represents false positives, and FN represents false negatives.

[0270] participate Figure 9 As shown, it is a flow chart of a method for identifying abnormal accounts in an embodiment of the present application. The process can be applied to a terminal device or a server. The specific process is as follows:

[0271] S901. Obtain at least one complaint text generated for the account to be detected within a set collection period.

[0272] The collection period can be set to one day, one week or one month, and there is no restriction on this, and it can be set according to actual recognition needs.

[0273] S902. Based on the target industry category to which the account to be detected belongs, a target recognition model is used to perform abnormal behavior analysis on at least one complaint text to obtain corresponding analysis results, wherein the target recognition model is obtained through iterative training using various historical complaint texts.

[0274] The analysis results include the prediction scores for each category label. The training process of the object recognition model is described above and will not be repeated here.

[0275] S903: Based on the at least one analysis result obtained, identify the account to be detected and obtain an account identification result of the account to be detected.

[0276] In one possible implementation, a complaint text generated for the account to be detected within a set collection period is obtained. Then, based on the target industry category to which the account to be detected belongs, a target recognition model is used to perform abnormal behavior analysis on the complaint text to obtain analysis results. Thereafter, based on the obtained analysis results, the account to be detected is identified to obtain an account identification result for the account to be detected.

[0277] Among them, the analysis results of the complaint text include the prediction scores corresponding to each category label. Based on the prediction scores corresponding to each category label and the set threshold, the category label with a value exceeding the set threshold is used as the final prediction result. Based on the prediction result, it is determined whether the account to be detected is an abnormal account.

[0278] As another possible implementation method, multiple complaint texts generated for the account to be detected within a set collection period are obtained. Then, based on the target industry category to which the account to be detected belongs, a target recognition model is used to perform abnormal behavior analysis on the multiple complaint texts to obtain corresponding analysis results. Thereafter, based on the analysis results of each of the multiple complaint texts obtained, the account to be detected is identified to obtain an account recognition result for the account to be detected.

[0279] Since a single complaint text cannot fully reflect the overall suspiciousness of an account, in an embodiment of the present application, the account identification result is obtained by comprehensively considering the abnormal behavior analysis results of multiple complaint texts. By aggregating the prediction results of a single complaint text to the account level, a comprehensive evidence-based abnormal probability assessment method is provided, thereby improving the accuracy of abnormality identification.

[0280] In some implementations, in S901, when the account to be detected meets the account detection conditions, at least one complaint text generated for the account to be detected is obtained according to the set collection cycle, and each complaint text is generated by the corresponding complaint account after resource transfer with the account to be detected.

[0281] Based on the at least one analysis result obtained, the account to be detected is identified. After obtaining the account identification result of the account to be detected, if the account identification result of the account to be detected indicates that the account to be detected has abnormal resource transfer behavior, an alarm is issued for the account to be detected.

[0282] Among them, the account detection condition can be periodic detection, or detection can be performed when the number of complaint texts received for the account to be detected exceeds the set text quantity threshold, or detection can be performed every time a complaint text is received for the account to be detected. There is no restriction on this.

[0283] Resource transfers include, but are not limited to, real resource transfers, virtual resource transfers, and virtual item transfers. For example, if account A transfers 10,000 yuan to account B, account A may file a complaint against account B; if account A transfers 200 game coins to account B, account A may file a complaint against account B; if account A gifted a certain game item to account B, account A may file a complaint against account B. In one possible implementation, when account A files a complaint against account B, it may submit relevant documentation related to the resource transfer. This documentation can then be used to filter out complaints related to resource transfers from a large volume of complaints.

[0284] In the embodiment of the present application, the alarm method includes but is not limited to: one or more of: telephone, text message, message push, etc.

[0285] In some implementations, when executing S903, the following methods may be used but are not limited to:

[0286] Method 1:

[0287] First, based on the value of the at least one predicted probability obtained, a predicted probability that meets a set evaluation condition is selected from the at least one predicted probability as a target probability;

[0288] Secondly, a weighted sum is obtained by taking the target probability and the average probability of at least one predicted probability and normalizing the weighted sum based on a set normalization parameter to obtain the abnormal probability of the account to be detected. The normalization parameter is determined based on the number of complaint texts.

[0289] Finally, based on the abnormal probability, the account identification result of the account to be detected is obtained.

[0290] During the collection cycle, the account to be tested may receive more than one complaint. Each complaint text corresponds to an analysis result. Therefore, if there are multiple complaint texts, the analysis results of multiple texts need to be aggregated as the account identification result of the account to be tested.

[0291] The account identification result may include the aggregation probability corresponding to each category label.

[0292] For example, for each category label, the following formula (9) can be used to calculate the corresponding aggregation probability:

[0293] P(suspicious=1|complaint)=(P max +P meam -(n / c)) / (1+(n / c)) Formula (9)

[0294] Among them, P(suspicious=1|complaint) represents the aggregate probability of a category label in each complaint sample, P max is the predicted probability with the highest value among all complaint texts. mean is the average predicted probability of each complaint text. n represents the number of complaint texts generated for the account under test. c is a constant used to adjust the sensitivity of probability aggregation. The value of c can be tuned on the training set using techniques such as cross-validation.

[0295] That is, the aggregate probability is calculated by summing the maximum probability and the average probability, and then dividing it by a normalization term (1+n / c). This normalization term increases with the increase in the number of complaint texts n, so that when there are more complaints, the impact of a single high-probability complaint text on the total probability is reduced.

[0296] The above implementation ensures that even when there is only one very high probability read, the overall probability will not be too low; the impact of other reads is incorporated by averaging the probabilities; and the probability is adjusted by the ratio of n and c. For accounts with many complaints, the overall probability will be normalized to avoid over-amplification.

[0297] In one possible implementation method, after obtaining the aggregate probability corresponding to each category label, the aggregate probability corresponding to each category label can also be used as the account feature of the account to be detected. Then, based on the account feature of the account to be detected, the anti-money laundering model used to identify accounts with specific abnormal behaviors is used to identify whether the account to be detected is suspected of specific abnormal behaviors, thereby improving the precision and recall rate of the anti-money laundering model in identifying abnormal accounts.

[0298] Method 2: Voting.

[0299] Specifically, for each complaint text in at least one complaint text, based on the analysis results corresponding to the complaint text, the complaint text is classified and predicted to determine the abnormal behavior category corresponding to the complaint text. Then, based on the abnormal behavior category corresponding to each of the at least one complaint texts, a voting strategy is set to determine the abnormal behavior category of the account to be detected.

[0300] The setting of the voting strategy includes but is not limited to: if the majority of complaint texts point to a specific abnormal behavior category, then the abnormal behavior category will be considered as the abnormal behavior category of the account to be detected.

[0301] In one possible implementation, each analysis result includes the predicted probabilities of each category label for the corresponding complaint text. For each complaint text, the category label whose predicted probability exceeds a set predicted probability threshold is used as the abnormal behavior category corresponding to the complaint text. Alternatively, each predicted probability can be used as a feature of the complaint text. Based on these features, an anti-money laundering model designed to identify specific crime categories can be used to identify the abnormal behavior category corresponding to the complaint text.

[0302] For example, there are 10 complaint samples for the account to be tested. Among the 10 complaint samples, the abnormal behavior category of 5 complaint samples is "gambling-online gambling", the abnormal behavior category of 3 complaint samples is "fraud-gambling fraud", and the abnormal behavior category of 2 complaint samples is "illegal fundraising-fund-raising fraud". Then, the abnormal behavior category of the account to be tested is "gambling-online gambling".

[0303] Method 3: Weighted average.

[0304] The predicted probabilities of individual complaints are aggregated using a weighted average approach, where weights can be assigned based on the severity of the complaint, its temporality (e.g., more recent complaints are weighted more heavily), or other relevant factors.

[0305] In one possible implementation, the text weight corresponding to each complaint text is evaluated separately from at least one evaluation dimension. Then, for each category label, based on the predicted probability of each complaint text in the category label and combined with the text weight corresponding to each complaint text, the aggregate probability of the account to be detected in the category label is obtained.

[0306] After obtaining the aggregate probability corresponding to each category label, the aggregate probability corresponding to each category label can be used as the account feature of the account to be detected. Then, based on the account feature of the account to be detected, the anti-money laundering model used to identify accounts of specific crimes can be used to identify whether the account to be detected is suspected of a specific crime, thereby improving the precision and recall rate of the anti-money laundering model in identifying abnormal accounts.

[0307] The evaluation dimensions include, but are not limited to, one or more of the severity of the complaint and the time of the complaint. The severity of the complaint can be determined based on the complaint level associated with the complaint text. For example, when evaluating the text weight, the higher the severity of the complaint, the greater the text weight, or the more recent the complaint time, the greater the text weight.

[0308] Method 4: Sequence model aggregation: Use sequence models such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to process chronological complaint texts to obtain merchant risk trends that change over time.

[0309] Exemplarily, after obtaining the prediction scores of each type of label corresponding to each complaint text, the prediction scores of each type of label corresponding to each complaint text are used as the prediction score sets corresponding to each complaint text, and the prediction score sets corresponding to each complaint text are sorted according to the complaint time corresponding to each complaint text, wherein each prediction score set contains the prediction scores of each type of label corresponding to the corresponding complaint sample. Then, the sequence model is used to process the sorted prediction score sets to obtain the account risk trend of the account to be detected that changes over time.

[0310] Method 5: Rule-based aggregation.

[0311] As a possible implementation method, according to the complaint level corresponding to the complaint text, the complaint text corresponding to the target complaint level is screened out from each complaint text, and then based on the screened complaint text, combined with method 1, method 2, method 3 or method 4, the account identification result of the account to be detected is obtained. Among them, the target complaint level can be one or more specified complaint levels, or it can be one or more complaint levels with the highest complaint severity in each complaint text. Of course, in actual application, the complaint sample screening is not limited to using the complaint level. The complaint sample screening can also be performed according to the set prediction probability threshold. There is no limitation on this and it will not be elaborated here.

[0312] Method 6: Probabilistic model aggregation. Considering the uncertainty of each complaint and the distribution of the overall data, a probabilistic model can also be used to aggregate the predicted probabilities of each complaint text to obtain the account identification results of the account to be detected. Among them, the probabilistic model includes but is not limited to the Bayesian method.

[0313] Method 7: Cluster Analysis: Cluster analysis is performed on the complaint text based on text features and predicted probabilities. The clustering results are used to identify the account to be tested. This account identification result can reveal the main problem areas that the account to be tested may have. In this embodiment of the application, the clustering method is not limited and will not be further described.

[0314] In some implementations, LLM may also be used to directly classify complaint texts and obtain analysis results corresponding to each complaint text.

[0315] The following describes the method in conjunction with a specific embodiment.

[0316] See Figure 10 As shown, this application involves the following stages: data preparation and sample annotation, model structure design, model training, model evaluation, and model application.

[0317] In the data preparation and sample labeling stage, first, obtain labeled initial samples; second, perform data enhancement on each labeled initial sample according to one or more of the following operations: data enhancement operation 1, data enhancement operation 2, and data enhancement operation 3, to obtain expanded samples; then, perform few-shot learning on the LLM based on the labeled initial samples to obtain the learned LLM; finally, use the learned LLM to perform multiple label predictions on each unlabeled initial sample to obtain the multiple label prediction results for each unlabeled initial sample, and obtain the soft label of each unlabeled initial sample based on the multiple label prediction results for each unlabeled initial sample. In addition, the category label of each labeled sample can also be converted into a soft label. Finally, a training sample set is constructed based on each labeled initial sample and each expanded sample. Each training sample consists of a historical complaint sample, industry type text, and a soft label.

[0318] During the model structure design phase, the initial recognition model was designed for multi-task learning. Specifically, Sentence-BERT was used as the encoder, a two-layer MLP was employed as the classification layer, and the Sentence-BERT embedding layer was used as the shared layer. In the recognition model, one Sentence-BERT was used to process the first-level labels, and one Sentence-BERT was used to process the second-level labels.

[0319] During the model training phase, first, a dynamic weighted loss is used as the loss weight of each sub-model. Second, the MSE loss function is used to calculate the loss weight of each sub-model. Next, the soft labels are processed. That is, for each sample data, the multiple classification results generated by the LLM are converted into a soft label vector. For example, if the classification results are 5 / 8, 1 / 8, and 2 / 8, this vector is directly used as the target output of the model. Finally, the pre-trained weights are loaded, and the pre-trained model is fine-tuned according to the above configuration.

[0320] During the model evaluation phase, evaluation metrics such as the F1 score are used to obtain a model evaluation value. When the model evaluation value meets the model evaluation conditions, the recognition sub-model for second-level label recognition in the trained initial recognition model is used as the target recognition model.

[0321] During the model application phase, based on the target industry category to which the account to be tested belongs, a target recognition model is used to perform abnormal behavior analysis on at least one complaint text to obtain corresponding analysis results. Then, based on at least one analysis result obtained, the account to be tested is identified to obtain an account recognition result for the account to be tested.

[0322] In an embodiment of the present application, the industry type is combined with the complaint text as part of the model input. This combination can help the model better understand the specific context of the complaint and specific industry-related issues. For example, in the model, the industry type and the complaint text can be treated as a pair of sentences, and the Sentence-BERT model can be used to extract features. This two-sentence input method can better capture the association between industry-specific semantic features and complaint texts. Complaint texts from different industries may contain industry-specific terms and expressions. The industry information provided can guide the model to more accurately identify complaint types related to specific industries, especially in issues unique to certain industries. Using industry text as auxiliary information can enhance the adaptability and sensitivity of the model when processing complaints from different industries.

[0323] In addition, a single complaint text cannot fully reflect the degree of abnormality of an account. Therefore, the prediction results of a single complaint text are aggregated to the account level, and by comprehensively considering the information of multiple complaints, an evidence-based abnormality assessment is provided.

[0324] Optionally, in multi-label classification tasks, different labels may have varying degrees of difficulty and importance. A traditional single loss function cannot effectively balance these differences. However, a multi-task learning framework and dynamic loss weighting optimize the model's performance across all labels. Furthermore, with a multi-task learning framework, each level of the task corresponds to a classification task, and dynamic loss weighting is used to optimize the model's overall performance across all tasks. This dynamic loss weighting method enables the model to automatically adjust its learning focus based on the performance of different tasks, improving overall classification accuracy.

[0325] Alternatively, compared to single-level labels, hierarchical and multi-label data often perform poorly, especially when sample distribution is uneven. Therefore, it may be beneficial to predict the label hierarchy by retaining high-level labels. By designing a multi-label classification method that adapts to the hierarchy, knowledge between label levels can be transferred. Even when samples of a certain sub-category are sparse, relevant text classification knowledge can still be learned through the upper-level labels.

[0326] Optionally, knowledge distillation using regression methods effectively transfers the LLM's knowledge to a small production model. This preserves some of the LLM's text understanding capabilities while reducing resource consumption, enabling the deployment of high-performance models on limited hardware. Leveraging the soft-label data generated by the LLM, multiple inferences are performed to obtain the soft label distribution for each sample. Through this distillation technique, the small production model can learn and mimic the LLM's predictive behavior. The soft label distribution obtained through multiple inferences provides the small model with a richer training signal, improving its performance and generalization on specific tasks.

[0327] Based on the same inventive concept, the present application embodiment provides an abnormal account identification device. Figure 11 As shown, it is a schematic diagram of the structure of the abnormal account identification device 1100, which may include:

[0328] The text acquisition unit 1101 is configured to acquire at least one complaint text generated for the account to be detected within a set collection period;

[0329] A text analysis unit 1102 is configured to perform abnormal behavior analysis on each of the at least one complaint texts based on the target industry category to which the account to be detected belongs, using a target recognition model to obtain corresponding analysis results, wherein the target recognition model is obtained through iterative training using various historical complaint texts;

[0330] The account analysis unit 1103 is configured to identify the account to be detected based on at least one analysis result obtained, and obtain an account identification result of the account to be detected.

[0331] In one possible implementation, the text analysis unit 1102 is further configured to:

[0332] Based on each historical complaint text, a training sample set is constructed. Each training sample carries corresponding label data. The label data is used to characterize the multiple abnormal behavior categories to which the corresponding training sample belongs. There is a hierarchical relationship between the multiple abnormal behavior categories.

[0333] Based on the training sample set, the initial recognition model is iteratively trained to obtain a trained initial recognition model, and based on the trained initial recognition model, the target recognition model is constructed.

[0334] In one possible implementation, the initial recognition model includes: multiple recognition sub-models, each recognition sub-model is used to predict a level of abnormal behavior category, and during each iteration, the text analysis unit 1102 is used to perform the following operations:

[0335] Obtaining a batch of training samples, and using the multiple recognition sub-models to perform label prediction on the training samples, to obtain prediction results corresponding to the training samples at multiple levels;

[0336] Based on the label data carried by each of the training samples and in combination with the multiple prediction results corresponding to each of the training samples, a model loss is obtained, and model parameters are adjusted based on the model loss.

[0337] In one possible implementation, when using the multiple recognition sub-models to perform label prediction on each training sample and obtaining prediction results corresponding to each training sample at multiple levels, the text analysis unit 1102 is specifically configured to:

[0338] For multiple recognition sub-models, perform the following operations respectively:

[0339] Using a feature extraction network in a recognition sub-model, feature extraction is performed on each of the training samples to obtain corresponding text features;

[0340] The classification network in the recognition sub-model is used to classify the obtained text features respectively, and obtain the prediction results corresponding to the training samples at the corresponding levels.

[0341] In one possible implementation, when constructing a training sample set based on the historical complaint texts, the text analysis unit 1102 is specifically configured to:

[0342] Based on the historical complaint texts and their corresponding industry type texts, construct initial samples, each initial sample including a historical complaint text and its corresponding industry type text;

[0343] According to the set data enhancement method, data enhancement is performed on the initial samples to obtain the expanded samples, and the training sample set is obtained based on the initial samples and the expanded samples.

[0344] In one possible implementation, when performing data enhancement on the initial samples according to the set data enhancement method to obtain the expanded samples, the text analysis unit 1102 is specifically configured to:

[0345] For at least one of the initial samples, perform at least one of the following operations:

[0346] Replacing at least one word contained in a historical complaint text in an initial sample with a synonym of the at least one word to obtain a corresponding expanded sample;

[0347] Replacing keywords contained in a historical complaint text in an initial sample with interference words similar in font to the keywords to obtain a corresponding expanded sample;

[0348] Based on the semantic information of the historical complaint text of an initial sample, similar text expansion is performed on an initial sample to obtain a corresponding expanded sample.

[0349] In a possible implementation, the text analysis unit 1102 is further configured to generate label data for a sample data in the following manner:

[0350] Using a label annotation model, a training sample is labeled multiple times to obtain labeling results corresponding to each of the multiple label annotations, each labeling result representing the abnormal behavior category to which the training sample belongs;

[0351] Results are aggregated based on the multiple labeling results obtained to obtain a soft label for the training sample, and the soft label is used as label data for the training sample, wherein the soft label includes a label probability that the training sample belongs to each abnormal behavior category.

[0352] In one possible implementation, based on the label data carried by each training sample and in combination with the obtained multiple prediction results corresponding to each training sample, when obtaining the model loss, the text analysis unit 1102 is specifically configured to:

[0353] Based on the label data carried by each of the training samples, combined with the obtained multiple prediction results corresponding to each of the training samples, obtain the sub-loss of each of the multiple recognition sub-models;

[0354] The model loss is obtained based on the sub-losses of the multiple recognition sub-models and the loss weights of the multiple recognition sub-models.

[0355] In one possible implementation, after obtaining the model loss based on the respective sub-losses of the multiple recognition sub-models and the respective loss weights of the multiple recognition sub-models, the text analysis unit 1102 is further configured to:

[0356] Based on the label data carried by each of the training samples, combined with the obtained multiple prediction results corresponding to each of the training samples, a model evaluation is performed to obtain a model evaluation value of each of the multiple recognition sub-models;

[0357] Based on the obtained model evaluation value, when determining to adjust the weight of at least one recognition sub-model among the multiple recognition sub-models, based on the loss weight of each of the at least one recognition sub-model in the current iteration process, the loss weight of each of the at least one recognition sub-model in the next iteration process is obtained.

[0358] In one possible implementation, when adjusting the model parameters based on the model loss, the text analysis unit 1102 is specifically configured to:

[0359] If the set parameter adjustment conditions are met, adjusting the other model parameters in the initial recognition model except for the model parameters whose parameter states are frozen based on the model loss, and adjusting the parameter states of the model parameters in the frozen state to a non-frozen state;

[0360] If the set parameter adjustment condition is not met, the other model parameters in the initial recognition model except for the model parameters whose parameter state is the frozen state are adjusted based on the model loss.

[0361] In a possible implementation, when constructing the target recognition model based on the trained initial recognition model, the text analysis unit 1102 is specifically configured to:

[0362] Based on the set target level, extracting the recognition sub-model corresponding to the target level from the trained initial recognition model;

[0363] A target recognition model is constructed based on the model parameters of the recognition sub-model corresponding to the target level.

[0364] In one possible implementation, each analysis result includes the predicted probability of each abnormal category of the corresponding complaint text;

[0365] Then, based on the obtained at least one analysis result, the account to be detected is identified. When the account identification result of the account to be detected is obtained, the account analysis unit 1103 is specifically configured to:

[0366] Based on the value of the at least one predicted probability obtained, selecting a predicted probability that meets the set evaluation conditions from the at least one predicted probability as a target probability;

[0367] Performing a weighted summation on the target probability and the average probability sum of the at least one predicted probability to obtain a weighted sum, and normalizing the weighted sum based on a set normalization parameter to obtain an abnormal probability of the account to be detected, wherein the normalization parameter is determined based on the number of complaint texts;

[0368] Based on the abnormal probability, an account identification result of the account to be detected is obtained.

[0369] In one possible implementation, when obtaining at least one complaint text generated for the account to be detected within a set collection period, the text obtaining unit 1101 is specifically configured to:

[0370] When the account to be detected meets the account detection conditions, at least one complaint text generated for the account to be detected is obtained according to the set collection period, where each complaint text is generated after the corresponding complaining account transfers resources with the account to be detected;

[0371] After identifying the account to be detected based on the at least one analysis result obtained and obtaining the account identification result of the account to be detected, the account analysis unit 1103 is further configured to:

[0372] If the account identification result of the account to be detected indicates that the account to be detected has abnormal resource transfer behavior, an alarm is issued for the account to be detected.

[0373] For the convenience of description, the above parts are divided into modules (or units) according to their functions and described separately. Of course, when implementing this application, the functions of each module (or unit) can be implemented in the same or multiple software or hardware.

[0374] Regarding the apparatus in the above embodiment, the specific manner in which each unit executes the request has been described in detail in the embodiment of the method, and will not be elaborated on here.

[0375] Those skilled in the art will appreciate that various aspects of the present application can be implemented as systems, methods, or program products. Therefore, various aspects of the present application can be specifically implemented in the following forms: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation that combines hardware and software aspects, which may be collectively referred to herein as a "circuit," "module," or "system."

[0376] Based on the same inventive concept, an embodiment of the present application further provides an electronic device. In one embodiment, the electronic device may be a server or a terminal device. Figure 12 , which is a schematic structural diagram of a possible electronic device provided in an embodiment of the present application, Figure 12 In the embodiment, the electronic device 1200 includes: a processor 1210 and a memory 1220.

[0377] The memory 1220 stores a computer program that can be executed by the processor 1210 . The processor 1210 can execute the steps of the above-mentioned abnormal account identification method by executing the instructions stored in the memory 1220 .

[0378] Memory 1220 may be a volatile memory, such as random-access memory (RAM); a non-volatile memory, such as read-only memory (ROM), flash memory, a hard disk drive (HDD), or a solid-state drive (SSD); or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 1220 may also be a combination of the above memories.

[0379] The processor 1210 may include one or more central processing units (CPUs) or digital processing units, etc. The processor 1210 is configured to implement the above-mentioned abnormal account identification method when executing the computer program stored in the memory 1220 .

[0380] In some embodiments, the processor 1210 and the memory 1220 may be implemented on the same chip. In some embodiments, they may also be implemented on separate chips.

[0381] In the embodiment of the present application, the specific connection medium between the processor 1210 and the memory 1220 is not limited. In the embodiment of the present application, the processor 1210 and the memory 1220 are connected via a bus as an example. Figure 12 The connections between the other components are shown in bold lines for illustration only and are not intended to be limiting. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of description, Figure 12 The diagram shows a single thick line, but this does not indicate that there is only one bus or one type of bus.

[0382] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, which includes a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to perform the steps of the above-mentioned abnormal account identification method. In some possible implementations, various aspects of the abnormal account identification method provided by the present application can also be implemented in the form of a program product, which includes a computer program. When the program product is run on an electronic device, the computer program is used to enable the electronic device to perform the steps in the above-mentioned abnormal account identification method. For example, the electronic device can perform the following steps: Figure 6 or Figure 9 Follow the steps shown in .

[0383] The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires, a portable disk, a hard disk, RAM, ROM, an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0384] The program product of the embodiments of the present application may be a CD-ROM and include a computer program, and may be run on an electronic device. However, the program product of the present application is not limited thereto. In this document, a readable storage medium may be any tangible medium that contains or stores a computer program that can be used by or in conjunction with a command execution system, apparatus, or device.

[0385] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries a readable computer program. Such a propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a computer program for use by or in conjunction with a command execution system, apparatus, or device.

[0386] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0387] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for identifying abnormal accounts, characterized in that: include: Obtain at least one complaint text generated for the account to be tested within the set collection period; Based on the target industry category to which the account to be detected belongs, a target recognition model is used to perform abnormal behavior analysis on each of the at least one complaint texts to obtain corresponding analysis results, wherein the target recognition model is obtained through iterative training using various historical complaint texts; Based on the at least one analysis result obtained, the account to be detected is identified to obtain an account identification result of the account to be detected.

2. The method according to claim 1, wherein Before obtaining at least one complaint text generated for the account to be detected within the set collection period, the method further includes: Based on each historical complaint text, a training sample set is constructed. Each training sample carries corresponding label data. The label data is used to characterize the multiple abnormal behavior categories to which the corresponding training sample belongs. There is a hierarchical relationship between the multiple abnormal behavior categories. Based on the training sample set, the initial recognition model is iteratively trained to obtain a trained initial recognition model, and based on the trained initial recognition model, the target recognition model is constructed.

3. The method according to claim 2, wherein The initial recognition model includes: multiple recognition sub-models, each recognition sub-model is used to predict a level of abnormal behavior category, and each iteration process performs the following operations: Obtaining a batch of training samples, and using the multiple recognition sub-models to perform label prediction on the training samples, to obtain prediction results corresponding to the training samples at multiple levels; Based on the label data carried by each of the training samples and in combination with the multiple prediction results corresponding to each of the training samples, a model loss is obtained, and model parameters are adjusted based on the model loss.

4. The method according to claim 3, wherein The method of using the multiple recognition sub-models to perform label prediction on each training sample to obtain prediction results corresponding to each training sample at multiple levels includes: For multiple recognition sub-models, perform the following operations respectively: Using a feature extraction network in a recognition sub-model, feature extraction is performed on each of the training samples to obtain corresponding text features; The classification network in the recognition sub-model is used to classify the obtained text features respectively, and obtain the prediction results corresponding to the training samples at the corresponding levels.

5. The method according to claim 2, 3 or 4, characterized in that The training sample set is constructed based on the historical complaint texts, including: Based on the historical complaint texts and their corresponding industry type texts, construct initial samples, each initial sample including a historical complaint text and its corresponding industry type text; According to the set data enhancement method, data enhancement is performed on the initial samples to obtain the expanded samples, and the training sample set is obtained based on the initial samples and the expanded samples.

6. The method according to claim 5, wherein The step of performing data enhancement on the initial samples according to the set data enhancement method to obtain the expanded samples includes: For at least one of the initial samples, perform at least one of the following operations: Replacing at least one word contained in a historical complaint text in an initial sample with a synonym of the at least one word to obtain a corresponding expanded sample; Replacing keywords contained in a historical complaint text in an initial sample with interference words similar in font to the keywords to obtain a corresponding expanded sample; Based on the semantic information of the historical complaint text of an initial sample, similar text expansion is performed on an initial sample to obtain a corresponding expanded sample.

7. The method according to claim 2, 3 or 4, characterized in that The label data of a sample data is generated in the following way: Using a label annotation model, a training sample is labeled multiple times to obtain labeling results corresponding to each of the multiple label annotations, each labeling result representing the abnormal behavior category to which the training sample belongs; Results are aggregated based on the multiple labeling results obtained to obtain a soft label for the training sample, and the soft label is used as label data for the training sample, wherein the soft label includes a label probability that the training sample belongs to each abnormal behavior category.

8. The method according to claim 3 or 4, wherein: Based on the label data carried by each training sample, combined with the obtained multiple prediction results corresponding to each training sample, the model loss is obtained, including: Based on the label data carried by each of the training samples, combined with the obtained multiple prediction results corresponding to each of the training samples, obtain the sub-loss of each of the multiple recognition sub-models; The model loss is obtained based on the sub-losses of the multiple recognition sub-models and the loss weights of the multiple recognition sub-models.

9. The method according to claim 8, wherein After obtaining the model loss based on the respective sub-losses of the multiple recognition sub-models and the respective loss weights of the multiple recognition sub-models, the method includes: Based on the label data carried by each of the training samples, combined with the obtained multiple prediction results corresponding to each of the training samples, a model evaluation is performed to obtain a model evaluation value of each of the multiple recognition sub-models; Based on the obtained model evaluation value, when determining to adjust the weight of at least one recognition sub-model among the multiple recognition sub-models, based on the loss weight of each of the at least one recognition sub-model in the current iteration process, the loss weight of each of the at least one recognition sub-model in the next iteration process is obtained.

10. The method according to claim 3 or 4, characterized in that The adjusting of model parameters based on the model loss includes: If the set parameter adjustment conditions are met, adjusting the other model parameters in the initial recognition model except for the model parameters whose parameter states are frozen based on the model loss, and adjusting the parameter states of the model parameters in the frozen state to a non-frozen state; If the set parameter adjustment condition is not met, the other model parameters in the initial recognition model except for the model parameters whose parameter state is the frozen state are adjusted based on the model loss.

11. The method according to claim 2, 3 or 4, characterized in that The step of constructing the target recognition model based on the trained initial recognition model includes: Based on the set target level, extracting the recognition sub-model corresponding to the target level from the trained initial recognition model; A target recognition model is constructed based on the model parameters of the recognition sub-model corresponding to the target level.

12. The method according to any one of claims 1 to 4, wherein Each analysis result contains the predicted probability of each abnormal category of the corresponding complaint text; Then, based on the obtained at least one analysis result, identifying the account to be detected and obtaining an account identification result of the account to be detected includes: Based on the value of the at least one predicted probability obtained, selecting a predicted probability that meets the set evaluation conditions from the at least one predicted probability as a target probability; Performing a weighted summation on the target probability and the average probability sum of the at least one predicted probability to obtain a weighted sum, and normalizing the weighted sum based on a set normalization parameter to obtain an abnormal probability of the account to be detected, wherein the normalization parameter is determined based on the number of complaint texts; Based on the abnormal probability, an account identification result of the account to be detected is obtained.

13. The method according to any one of claims 1 to 4, wherein The obtaining of at least one complaint text generated for the account to be detected within the set collection period includes: When the account to be detected meets the account detection conditions, at least one complaint text generated for the account to be detected is obtained according to the set collection period, where each complaint text is generated after the corresponding complaining account transfers resources with the account to be detected; After identifying the account to be detected based on the at least one analysis result obtained and obtaining the account identification result of the account to be detected, the method further includes: If the account identification result of the account to be detected indicates that the account to be detected has abnormal resource transfer behavior, an alarm is issued for the account to be detected.

14. An abnormal account identification device, characterized in that: include: A text acquisition unit, configured to acquire at least one complaint text generated for the account to be detected within a set collection period; a text analysis unit configured to perform abnormal behavior analysis on each of the at least one complaint texts based on the target industry category to which the account to be detected belongs, using a target recognition model to obtain corresponding analysis results, wherein the target recognition model is obtained through iterative training using various historical complaint texts; The account analysis unit is configured to identify the account to be detected based on at least one analysis result obtained, and obtain an account identification result of the account to be detected.

15. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is enabled to perform the steps of any one of the methods of claims 1 to 13.

16. A computer-readable storage medium, characterized in that The method comprises a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to execute the steps of any one of the methods according to claims 1 to 13.

17. A computer program product, characterized in that It includes a computer program, which is stored in a computer-readable storage medium. A processor of an electronic device reads and executes the computer program from the computer-readable storage medium, so that the electronic device performs the steps of any method described in claims 1 to 13.

Images